GCS ranged requests sometimes fails with 400

Summary

In GitLab.com, we recently ran into an incident where a small number of Pages hosts failed with 500 errors. Digging into it we found that this resulted from internal error where ranged requests to GCS were failing with "400 Bad Request".

This behavior seems recent (started Sepetmeber 8th), judging by graphs from GCS: .

Kibana logs doesn't provide much info beside the error itself, so perhaps we may need to consider logging the response body of 400 responses.